Portuguese NER- TempClinBr - BioBERTpt(bio)
Treinado com BioBERTpt(bio), com o corpus TempClinBr.
Metricas:
precision recall f1-score support
0 0.44 0.29 0.35 28
1 0.75 0.60 0.66 420
2 0.57 0.40 0.47 10
3 0.57 0.36 0.44 11
4 0.70 0.85 0.77 124
5 0.72 0.67 0.69 291
6 0.84 0.90 0.87 2236
7 0.78 0.77 0.77 112
8 0.85 0.75 0.80 503
9 0.64 0.56 0.60 78
10 0.81 0.82 0.81 71
11 0.82 1.00 0.90 33
accuracy 0.81 3917
macro avg 0.71 0.66 0.68 3917
weighted avg 0.81 0.81 0.80 3917
Parâmetros:
device = cuda (Colab)
nclasses = len(tag2id)
nepochs = 50 => parou na 16
batch_size = 16
batch_status = 32
learning_rate = 3e-5
early_stop = 5
max_length = 256
write_path = 'model'
Eval no conjunto de teste - TempClinBr OBS: Avaliação com tag "O" (label 7), se necessário fazer a média sem essa tag.
tag2id ={'I-Ocorrencia': 0,
'I-Problema': 1,
'I-DepartamentoClinico': 2,
'B-DepartamentoClinico': 3,
'B-Ocorrencia': 4,
'B-Tratamento': 5,
'O': 6,
'B-Teste': 7,
'B-Problema': 8,
'I-Tratamento': 9,
'B-Evidencia': 10,
'I-Teste': 11,
'<pad>': 12}
precision recall f1-score support
0 0.59 0.20 0.29 51
1 0.77 0.69 0.73 645
2 0.67 0.71 0.69 14
3 0.87 0.43 0.58 30
4 0.71 0.80 0.75 146
5 0.79 0.77 0.78 261
6 0.84 0.93 0.88 2431
7 0.80 0.66 0.73 194
8 0.87 0.83 0.85 713
9 0.83 0.62 0.71 146
10 0.98 0.91 0.94 128
11 0.54 0.21 0.30 99
accuracy 0.83 4858
macro avg 0.77 0.65 0.69 4858
weighted avg 0.82 0.83 0.82 4858
Como citar: em breve
